R. Kamimura and K. Aoyama (Japan)
In this paper, we propose a new method to enhance hidden unit activations of multi-layered networks and to retrain networks imitating enhanced hidden unit activations. One of the most serious problems of multilayered networks consists in difficulty in interpreting final internal representations obtained by learning. Though many methods have been proposed to solve this problem, the problem still remains unsolved. For this problem, we propose a method to enhance hidden unit activations with explicit responses to input patterns. In addition, we propose a new method to retrain networks by imitating the enhanced hidden unit activations. We applied the method to a student survey, a cabinet approval rating estimation and another student survey. In all these problems, we succeeded in producing explicit and interpretable final representations.
Important Links:
Go Back
